I was going to suggest basically what you said in the first paragraph, but I didn't want to influence your own creativity =D

As for the second paragraph... depending on exactly what you mean by machine learning... say, possibly a more brute force, genetic-mutation, run 100s or 1000s of iterations to generate a 'strategy' ... that may be overkill.

I would suggest looking into Goal Oriented Action Planning, GOAP. GOAP is basically a theory, a conceptual framework of how to program AI, in the video game sense of AI... but it is way, way more computational lightweight.

It is also what made FEAR's AI so distinct, (the FEAR AI programmer invented it and named, has either an academically published or at least technical 'white paper' on it), convincing players that it was actually intelligently reacting to them, the environment, and making its own decisions... because arguably, it was, or at least doing a much more 'realistic' approximation of this at a fundamental level.

The difference between this game, your game's implementation of GOAP and most other ones... would be that it is like, the levels themselves, the sort of plot line itself, the 'dungeon master'... that is the Goal Oriented thinking agent, not just an npc entity in the world.

You can find some good explanations of how it works, how to implement it in a game engine, on youtube... though ironically I must warn you that a lot of tutorials from basicslly indie devs... they often do not fully even grasp the concept.

Which is funny to me, as they are not intelligent enough to understand a model of machine intelligence, yet feel confident enough to explain it or criticize it, lol.

But yeah, if you have both a comprehensive tracking and logging system for the player's actions... well, oversimplified, GOAP roughly invovles a sort of pathfinding algo traversing a node graph, with weights between nodes... and connections between nodes, or the 'cost' of traversing them... changing based on the agent's current situation.

So... you would roughly be trying to do a 'best fit approximation' of the player's own decision making process... into the the node graph the game itself uses... and then you could maybe reproduce the player themselve's approach to decision making, and then make that into an npc agent, if you wanted them to truly like, mirror fight someone that 'thinks' like them, at some point.